Today on The Anvil: Google's I/O reframes the developer stack around agents, Cursor answers with a cheaper Composer 2.5 (and a SpaceXAI training deal), and Iran's war hits day 82 with the Senate moving a War Powers rebuke and Tehran floating internet-cable tolls for Google and Meta. Plus Spokane locks in a 20-year growth map and Orange County's board meeting goes off the rails.
Google used I/O 2026 to ship Gemini 3.5 Flash (claimed 4x faster than competitor frontier models at sub-half pricing), Gemini Omni Flash for multimodal video generation with conversational editing, Gemini Spark as a persistent 24/7 personal agent, and Antigravity 2.0 β a standalone agent-first desktop app with dynamic subagents, parallel task execution, JSON hooks, and cron-style scheduling. AI Studio now generates Kotlin/Jetpack Compose Android apps end-to-end with in-browser emulator and Play Console integration. Pichai disclosed $180β190B capex and 7x YoY token volume growth to 3.2 quadrillion monthly.
Why it matters
Google is no longer chasing Anthropic and OpenAI on model quality alone β it's pricing the agent layer to be the default. Antigravity 2.0 as a standalone app (not an IDE plugin) is the structural tell: Google believes the editor surface itself is changing, and a $100 Ultra tier slots between Pro and the $200 flagship to capture professional builders. For anyone evaluating a coding-tool stack, the question is no longer 'which model is smartest' but 'whose orchestration layer holds up at scale,' and Google just put a credible third option on the table.
Yesterday's briefing covered Composer 2.5's benchmark numbers and pricing. The new development today: Cursor disclosed Composer 2.5 is built on Moonshot's open-source Kimi K2.5 base (raising China-origin model questions for enterprise/government buyers) and simultaneously announced a SpaceXAI partnership to train a significantly larger from-scratch model using 10x compute on the Colossus 2 cluster (~1M H100-equivalent GPUs). The targeted-textual-feedback RL technique β correction hints at specific error points rather than only final reward β is what made the long-horizon multi-file performance work.
Why it matters
Cursor is pivoting from IDE wrapper to vertically integrated model company, and the SpaceXAI compute deal is the strategic mirror of Anthropic's SpaceX deal from two weeks ago β the coding-tool layer is now competing for hyperscaler compute access alongside the foundation labs. The Kimi K2.5 dependency is the wrinkle: it's the cheapest path to frontier coding quality but creates a provenance question for regulated buyers that Antigravity 2.0 doesn't have. Watch whether enterprise procurement starts asking about base-model lineage the way it asks about hosting region today.
Following yesterday's news that Anthropic will brief G20 finance ministers on Mythos cybersecurity findings, the company revised Project Glasswing rules to let its 40β50 partner organizations share Mythos-discovered vulnerabilities with external security teams, industry bodies, regulators, open-source maintainers, and media under responsible-disclosure norms. Mythos has reportedly identified thousands of zero-days with an 83% working-exploit success rate.
Why it matters
The structural asymmetry remains β Glasswing members still see findings first β but the policy shift acknowledges that defender advantage at 40β50 organizations is too narrow to be socially defensible when the underlying capability touches 'every major OS and browser.' This is the model for how frontier-AI capability disclosure gets normalized: tight initial circle, regulatory pressure, gradual widening. Worth watching as a template for what happens next with biosecurity and autonomous-cyber capabilities.
OpenAI formally joined the C2PA provenance standard and partnered with Google to embed invisible SynthID watermarks in its AI-generated images, with a public verification tool now in preview (currently limited to OpenAI outputs). Separately, Google's new Gemini Omni Flash video model ships with SynthID watermarking on by default and deliberately withholds speech-editing capability pending responsible-use assessment.
Why it matters
Provenance has quietly moved from policy paper to shipping default. The dual-layer approach (visible C2PA metadata + invisible SynthID) creates redundancy against stripping, and OpenAI joining a standard it long held out on is the signal β the AI-content-authenticity stack is consolidating around a small set of interoperable primitives. For anyone building products that ingest or display third-party media, the question shifts from 'should we verify provenance' to 'which libraries do we depend on.'
Orange County Supervisors Janet Nguyen and Don Wagner publicly attacked Vice Chair Katrina Foley on May 19 over her unilateral announcement of an herbicide-spray pause, with Nguyen comparing Foley to a corrupt former supervisor. Chairman Doug Chaffee gaveled a 10-minute break to cool the room. The dispute centers on individual supervisor authority versus board approval for staff direction and media communications.
Why it matters
The on-camera meltdown is the surface; the substance is a real governance question about how much unilateral authority an individual supervisor has to direct county staff or pause programs. Foley is the same supervisor who funded the San Clemente helopod from her office budget β a pattern of acting first, briefing the board later. Expect a formal rules-of-procedure fight at the next meeting and likely campaign-cycle fallout.
Accenture made a strategic investment in Aera Technology, whose platform deploys autonomous AI agents that monitor changes and execute actions across supply chain, procurement, finance, and operations. Hershey is cited as an early production customer. The investment lands the same week MondelΔz announced AI/automation rollouts across up to five distribution centers, Romark deployed Dexory's autonomous cycle-count robots, Mecalux scaled compute infrastructure to host agents inside its software suite, and Locus Robotics acquired Nexera for AI-driven mobile manipulation.
Why it matters
The center of gravity in supply-chain AI has shifted from 'planning tools that recommend' to 'agents that execute,' and Accenture putting capital behind Aera is the consultant-side signal that this is the integration story for 2026. Combined with this week's Blue Yonder + NVIDIA training-factory news and Palantir's forward-deployed-engineer model, the competitive matrix now has three distinct shapes: platform agents (Aera, Blue Yonder), custom builds (Palantir), and embedded vendor agents (Mecalux). The differentiator isn't model quality β it's data integration and the willingness to put agents in execution paths.
Uber's engineering team published the architecture behind DeepETT, a deep-learning traffic-forecasting system that improves long-trip ETA accuracy by 6%, increases forecast variance explained by 19%, and serves 2 million real-time forecasts per second. The system uses a transformer with fixed-size multi-view inputs to avoid dynamic graph computation at scale, separates segment-level prediction from trip-level routing, and runs continuous calibration to prevent drift. Estimated incremental impact: ~$100M annual revenue.
Why it matters
This is the rare production-systems write-up that explains how to make graph-shaped problems work at hyperscale without the graph: by linearizing the inputs, separating prediction from optimization, and treating drift as a first-class operational concern. The pattern is directly transferable to any logistics or route-optimization product where the naive approach is 'just put it in a GNN' β and the calibration approach is the part most production teams underinvest in.
1Password published a detailed case study of building an agent-driven pipeline from Jira ticket to design-system PR, by exposing their Knox design system over MCP and encoding implicit team conventions as narrowly-scoped skills. The post outlines six adaptable principles: narrow skill scope, human qualification gates, uncertainty flagging, explicit context handoffs, design-token-aware tooling, and a design-to-prototype spinoff that generates working prototypes from intent.
Why it matters
This is the most concrete production case study of the 'components as agent contracts' thesis β the same architecture Design.MD launched on Product Hunt and that Google Stitch and Claude Code use via DESIGN.md. The MCP-exposed design system converts static documentation into a queryable, executable surface; the narrowly-scoped skills are the machine-readable constraint layer that keeps agents from drifting into Tailwind defaults. For anyone running a design system, the 1Password post is the closest thing yet to a migration blueprint: which conventions are tacit enough that agents will fail without explicit skills is the question to bring to your next planning meeting.
A new Chrome extension, Markagent, converts UI element clicks into structured markdown prompts containing component names, file paths, CSS selectors, screenshots, and context for direct export to Claude Code, Cursor, Codex, and other AI assistants. Supports a journey mode for multi-step interactions and runs fully local with no telemetry.
Why it matters
The bottleneck in design-to-code feedback loops has shifted from 'can the agent fix the bug' to 'can I describe the bug fast enough to be worth using an agent at all.' Markagent attacks exactly that friction β the cognitive overhead of translating a visual observation into a structured, selector-aware prompt. For anyone working at the design-code seam, this is the kind of small tool that disproportionately changes daily workflow.
On May 18, Spokane City Council voted 6-1 to adopt a hybrid 'Preferred Alternative Growth Map' combining Alternatives 2 and 3 from the Plan Spokane 2046 process β projecting 20,000 new residents by 2046 and focusing development along transit corridors and near downtown, Hillyard, and west-plains employment hubs. The resolution guides land-use decisions but does not immediately rezone or authorize projects. Councilman Cathcart cast the lone dissent, flagging incomplete underlying planning data β consistent with his recent push for council oversight of road conversions.
Why it matters
This is the 20-year framework that will shape every infrastructure, transit, and zoning fight in Spokane for the next two decades. Transit-oriented + employment-hub focus aligns with how the city actually grew during the post-pandemic in-migration, but the gap between a guiding map and the parcel-level zoning code is where the next round of NIMBY battles will play out. Cathcart's dissent β and his parallel push for council authority over Transportation Commission decisions β is the pattern to watch.
Three new developments since yesterday's postponed strike. (1) Senate advanced a War Powers Act resolution 50-47 to halt US military operations in Iran β Sen. Bill Cassidy (R-LA) providing the decisive margin, the first procedural win for war-powers opponents, driven partly by gas prices over $4.53/gal and the expired 60-day authorization window. (2) Trump set a 'two-to-three-day' deadline for Iran to reach a deal while VP Vance described talks as in a 'pretty good' place. (3) Iran's Revolutionary Guard explicitly threatened to extend the war 'beyond the region' with 'crushing blows' if strikes resume. Foreign Policy analysis frames the underlying gap: Iran's 5-year moratorium offer versus the US 20-year ask on enrichment β with zero enrichment as a stated demand assessed as unattainable.
Why it matters
The Senate vote is the new signal. Republican fracture on continued operations β even against a likely veto β materially raises the political cost of escalation and gives Gulf-state mediators (Qatar, Saudi, UAE, who requested Tuesday's pause) a domestic-politics argument to press. The IRGC threat to extend 'beyond the region' is the behavioral counterpoint to diplomatic optics; read alongside the PGSA institutionalization and the ceasefire-window repositioning of ballistic missile assets from yesterday, the gap between diplomatic tone and military posture is widening, not narrowing. Watch the Friday-Saturday window for whether the 'two-to-three-day' clock produces a framework or a resumption.
TIME details the cable-toll dimension of the Persian Gulf Strait Authority (covered in yesterday's briefing), fleshing out the proposal for Iran to charge tech companies β Google, Meta, Microsoft β annual fees for fiber-optic cables crossing the Hormuz seabed. The cables carry the bulk of intercontinental data linking Europe, Asia, and the Gulf; coordinated disruption could create simultaneous shocks to energy and digital infrastructure affecting roughly $10 trillion in daily financial transactions.
Why it matters
The institutional move from 'closure threats' to 'permit regimes and tolls' is the real escalation β it tries to make leverage durable rather than episodic. For anyone whose product depends on intercontinental routing (which is most cloud workloads), this is a reminder that the Hormuz chokepoint is now a digital chokepoint, not just an oil one. Diversified routing through alternative cable paths is the multi-year structural response; the near-term response is monitoring.
Babel Street launched Insights Investigator, an agentic OSINT capability where analysts state intent in natural language and AI agents execute approved multi-step workflows across Babel Street's multilingual, rights-cleared data foundation. Early users report 50%+ reductions in research time. The launch lands alongside War on the Rocks' Salt Typhoon analysis arguing China's data-centric intelligence strategy is creating a structural decision-speed advantage over US agencies constrained by tighter legal guardrails.
Why it matters
Agentic OSINT is following the same arc as agentic coding: from 'demos of single tasks' to 'analyst-in-the-loop dispatch of multi-step investigations.' The War on the Rocks framing is the strategic backdrop β the question is no longer who has the best tools, but who can fuse heterogeneous open and commercial data fastest. The legal-guardrail gap is real and not closing, which means the US response increasingly looks like 'buy from commercial OSINT vendors' rather than build in-house.
The IDE is becoming a dispatcher Antigravity 2.0, Composer 2.5, AI Studio's Android target, and Markagent all push the same direction: the editor surface is now where you delegate, monitor, and review work β not where you type it. Multi-agent orchestration and parallel subagents are the new defaults.
Coding-tool economics are diverging from frontier model economics Cursor's Composer 2.5 on a Kimi K2.5 base hits near-Opus quality at ~1/10 token cost; Gemini 3.5 Flash advertises 4x faster at half-price. The tools that win in 2026 won't be the ones with the smartest single model β they'll be the ones that orchestrate cheap, fast agents in parallel.
Iran negotiations are running on Trump's compressed clock Strike postponed Tuesday, Senate War Powers vote Wednesday, 'two-to-three-day' deadline Thursday, Revolutionary Guard threatening to extend the war 'beyond the region.' The signal-to-noise ratio is collapsing; the underlying enrichment-rights impasse hasn't moved.
Supply chain AI is past the pilot phase Rockwell's 2026 survey: 34% of operations already AI-augmented, 90% calling digital transformation essential. MondelΔz, Romark, Mecalux, Locus/Nexera, Accenture+Aera all shipped concrete deployments this week β but only 43% of collected data is being used effectively, which is where the next round of value lives.
Provenance and watermarking are quietly becoming table stakes OpenAI joined C2PA and is co-watermarking with Google's SynthID. Gemini Omni Flash ships with SynthID on by default and withholds speech editing pending review. The 'responsible AI defaults' conversation has moved from white papers into product shipping decisions.
What to Expect
2026-05-22—Trump's stated 'two-to-three-day' deadline for an Iran deal expires; renewed strike risk.
2026-05-26—Laguna Beach shade-structure ban takes effect on 95% of city beaches.
2026-05-29—Israel-Lebanon security talks under the 45-day ceasefire extension.
2026-06-01—Huntington Beach's $50K/month housing-element penalties begin compounding.
2026-06-01—GitHub Copilot's AI Credits consumption pricing transition.
How We Built This Briefing
Every story, researched.
Every story verified across multiple sources before publication.
🔍
Scanned
Across multiple search engines and news databases
782
📖
Read in full
Every article opened, read, and evaluated
157
⭐
Published today
Ranked by importance and verified across sources
13
β The Anvil
π Listen as a podcast
Subscribe in your favorite podcast app to get each new briefing delivered automatically as audio.
Apple Podcasts
Library tab β β’β’β’ menu β Follow a Show by URL β paste